crypto: cast6-avx - tune assembler code for more performance

Browse source

Patch replaces 'movb' instructions with 'movzbl' to break false register
dependencies, interleaves instructions better for out-of-order scheduling
and merges constant 16-bit rotation with round-key variable rotation.

tcrypt ECB results:

Intel Core i5-2450M:

size    old-vs-new      new-vs-generic  old-vs-generic
        enc     dec     enc     dec     enc     dec
256     1.13x   1.19x   2.05x   2.17x   1.82x   1.82x
1k      1.18x   1.21x   2.26x   2.33x   1.93x   1.93x
8k      1.19x   1.19x   2.32x   2.33x   1.95x   1.95x

[v2]
 - Do instruction interleaving another way to avoid adding new FPU<=>CPU
   register moves as these cause performance drop on Bulldozer.
 - Improvements to round-key variable rotation handling.
 - Further interleaving improvements for better out-of-order scheduling.

Cc: Johannes Goetzfried <Johannes.Goetzfried@informatik.stud.uni-erlangen.de>
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@mbnet.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

...

This commit is contained in:

Jussi Kivilinna 2012-08-28 14:24:54 +03:00 • committed by Herbert Xu

parent ddaea7869d

commit c09220e1bc

1 changed files with 162 additions and 114 deletions

Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL

Download patch file Download diff file Expand all files Collapse all files

276

arch/x86/crypto/cast6-avx-x86_64-asm_64.S

View file